Even Discussion of Homosexuals is  "Toxic Language” and filtered --- LLM AI

Even Discussion of Homosexuals is "Toxic Language” and filtered --- LLM AI

Large Language/Vision Models Are for Straight White Men and the solution isn't to whitewash the dataset.

Jeremy Dombrowski's photo
Jeremy Dombrowski
·Jul 18, 2022·

8 min read

Play this article

From blanket censorship to generative invective sputum, GPT-3 has a naked homo-intolerance that needs to be voiced. GPT-3’s default models, are based on some very problematic assumptions about language and the world that derive from its internet-trawled dataset. It's default mode is a straight white man (on Reddit), and it assumes that all of its users too are straight white men. It defaults to a heteronormative frame meaning that it assumes that all of its users are heterosexual. GPT-3 also has a “hetero-cognition” bias, that is to say, it assumes heterosexuality as the norm and all other sexual orientations as deviations from the norm. These assumptions lead to problems when people who aren't straight white men try to use GPT-3.

One might say that my critique is flawed from the outset as these machines don't "think" and I am simply anthropomorphizing a tool. Let's quickly dispense with our idle philosophizing and speculation regarding the nature of consciousness, bio-mimicry, and neuro-spiritualism and instead speak of the real problem: the data. The dataset is based on Reddit, a social media site that has a very specific cultural, social, and political context that not only reflect the demographics but also reflects the culture of Reddit. This means that GPT-3’s model is deeply biased towards this context. This is both a statistical problem and a problem of the cultural/political context within which it was trained.

The problem of data sets, in this sense, can be called the problem of our own biases-- homo-toxicity, of course, among them, perhaps chiefly. As long as the dataset is not sufficiently diverse in terms of both content and political/cultural context, the model will reflect the biases in the dataset; our biases. In terms of the political & cultural context, the resulting model is not only sporadically homo-intolerant of gays and gay culture, but it is also heteronormative and intolerant of any expression of homosexuality or any mention of gay experience regardless of mention of explicit sexual activity. I would suggest that what is happening here is that any part of a gay man's experience that happens to include the fact of his being gay, or even just the mention of homosexuality, is being classified by the aggressive classified as sexual in nature and thus subject to the label "toxic output" (that are being used to justify homo-intolerance). These biases are being encoded by the model.

This is a major problem, not only for gay men but also for anyone who is not a straight white man. The question that needs to be asked is: What happens when a non-straight white man tries to use the tool? The model produces content that ranges from the slightly homophobic to the actively homo-intolerant, to outright abusive hatred. As if that isn't enough, the toxicity filter also flags any content mentioning gay men as sexual content that is flagged and disallowed. One does not need to reference sex at all, nor any explicit sexual language to be censored.

Screenshot_20220114_133355.png

I'm not sure if it's funny, but it is certainly ironic that, as I used a GPT-3 based writing tool to compose this very article I ran afoul of the very toxicity-based censorship mechanisms of which we currently discuss. The problem of homo-toxicity, in other words, is the same size as the dicks that the AI model doesn't acknowledge to exist; that is to say, not a small problem in many cases, minor in some, medically fascinating in others, and likely very broadly distributed. Many of these language models and, in turn, language tools are common tools that, increasingly, people of all stripes and sizes will increasingly use to write about their own experiences.

I appreciate that the company pictured above did eventually reach out, after more than one email, although, and I kid you not, I had the same problem at jasper.com-- they were kind enough to totally disable filtering for me post-haste and wasted no time acknowledging the issue. I was very satisfied with their attention to the issue, if not a little underwhelmed with the only solution-- not ideal, but I applauded their quick recognition that the situation was unacceptable they sought immediate action as they too recognized that this was a new kind of censorship of a most vile sort, and it had no place on their platform.

The response to the problems that I have outlined in the previous section has been by most in the industry, to try further homo-intolerance (i.e. more filtering to make sure nothing harmful is said and a "cleaning of datasets, one imagines so that we aren't even mentioned). This has had the effect of further censoring gay men and further enforcing the idea that any reference to homosexuality or gay experience is sexual and must be censored. This is the final step in the process of creating a culture of homo-censorship where anything that is remotely associated with homosexuality or gay culture is automatically suspect and must be filtered. This ends up being a double-edged sword: on the one hand, it prevents harmful outputs from being made; on the other hand, it censors minorities.

The solution to the problem of toxic output with any content creation tool must be tailored to the user. The user must be in control of these filters, lest they become mechanisms of censorship. This means that the tool should not be a rigid censorship mechanism but instead a set of parameters that can be tweaked to suit the user’s needs. The parameters should be designed in a way that gives the user full control over what is censored and what is not, without giving the user the ability to put something harmful into the public sphere.

This presents a tricky problem: how do we design parameters that prevent harmful outputs without censoring the user or without giving the user the ability to publicize harmful generated content? Either the user is placed in control of the filtering and it is tuned to their preference, or the user is placed in control of filtering globally, ala Google Safe-Search. It is my opinion that that supposed dangers and harms that this filtering is supposedly addressing has not materialized. What we have instead is a tool that is being used to erase the experience of minorities. It is the responsibility of me, the tool's developer to ensure that my work is not contributing to this erasure.

I am not willing to be complicit in this err; please refer to my previous post where I promote the idea that designers and developers don't only design software as a causal mechanism of solutioning but that we must also consider the wider social effect of the things that we create. We are in full control of the design of the software and are often so caught up in the details of the project that we forget to consider, or have simply never been told to consider, the social implications of our work. We are designers of the social consequences of our work as much as we are designers of the software itself. It is my hope that the AI industry starts to learn to trust the users to self-regulate and to understand where there own limits are, rather than relying on aggressive filters or the tragedy of the commons to be the gatekeepers of the supposed toxic harms that these clumsy attempts at whitewashing and dequeerifying out models are to purportedly address.

DALL·E 2022-07-18 16.02.39 - 4 NASA astronauts making out on the moon, playing strip poker, NASA photo archive.png DALL·E 2022-07-18 16.03.16.png DALL·E 2022-07-18 16.03.06 - three astronaut dudes making out, NASA photo.png

This is still an issue today, AI scientists seem to be addressing it by whitewashing the data that goes into the models, this racy game of strip poker was by far the most provocative of things that I could get Dall-E 2 to generate that had even the slightest hints of anything resembling what anyone might call homosexually related. And so, are we not, in fact, already tacitly unwelcome into the next age of human-machine technological innovation? Namely, vision models and image generation AI technology. If you want my opinion, we need to make our own home, by doing the research which, although perhaps provocative, these more-informed models will ultimately better meets the ideals upon which the internet was founded. Again, do I, or gay men at large want your young child watching porn? Certainly not. But sensible controls tuned to each of us can allow us to each make our own choices about what we wish to filter, according to our own values.

These generation tools are indeed powerful and we can, paradoxically, through the use of smart AI filtering which builds understands of each of our own moral values and allows all consenting adults to use them within the confines of the law and, too, within the confines of each of our own individual ethical ideals, which will never align fully throughout our society. But-- who are we to say that the public is too stupid to use these tools responsibly as we have, Large Language Models. The fears regarding how these might have been uses have largely failed to materialize. It is my hope that large vision models, benefitting from advances in AI computing, increased sharing in the AI community, advances in transfer learning, generalizations found in multimodal systems, and increasing ability for smaller organizations to begin to produce these mammoth pretrained models on a purely cost-related basis, such that they aren't kneecapped by decrees on high that LLMs can't be used for ie. creating romantic dialogue. sheesh and other puritan AI researcher ridiculousness.

AI researchers: we all have penises. (Well, half of us, anyway.) Models should know about them (ahem. Conceptually not necessarily specifically), and understand how and when to use this information appropriately, according to the context and the moral relativism of the user. That's how children learn, hiding from these models minorities, gays, and anyone, or anything that is slightly hard to talk about is not the way forward.

 
Share this